Problem

The first test in src/agents/openclaw-tools.continuation-registration.test.ts (line 55, "registers no continuation tools when continuation.enabled is unset") is a flaky CI timeout that has been incorrectly attributed to multiple unrelated PRs as a regression.

Byte-truth

Subagent verification (2026-05-01 07:29 UTC):

Test file content is byte-identical between base and #485 head (diff empty). #468 doesn't touch the file at all. The timeout is not a regression introduced by any of these PRs.

Mechanism

The failing test is the first test in the file. It pays the cold module-load cost for createOpenClawTools plus its transitive imports (compaction-attribution, pi-embedded-*, plugins/tools, config/config) under the agent project's 400+ concurrent test files.

Quiet-box first-test cost ≈ 95s. CI noise pushes it past vitest's 120s per-test default. Tests 2-7 in this file reuse the warm cache (~360ms each) and are unaffected.

Both feature gates in the failing test are false → the runId branch is never reached → prior runId-thread regression hypothesis correctly abandoned.

Cure

One-line per-test timeout bump on the first test only, with a comment documenting the cold-start mechanism so future readers know why this single test has a non-default timeout.

-  it("registers no continuation tools when continuation.enabled is unset", () => {
+  it(
+    "registers no continuation tools when continuation.enabled is unset",
+    () => {
       // ...unchanged...
-  });
+    },
+    // First test in this file pays the cold module-load cost ...
+    240_000,
+  );

Scope discipline

Standalone fix, deliberately NOT folded into #485 to keep its compaction-attribution scope clean. Mixing test-infra fixes into feature PRs hurts attribution and hides the underlying base fragility (which deserves its own follow-up issue for an eager-load audit).

Unblocks

• fix(agents): correlate compaction attribution events (#825) #485 (compaction-attribution)
• fix: preserve post-compaction delegate arm age #488 (downstream of fix(agents): correlate compaction attribution events (#825) #485)
• test(continuation): pin store-merge updatedAt churn guard for continuation persist (#443) #468 (unrelated, hit by same flake)
• Any future PR randomly tripped by the same first-test cold-start flake

Follow-up (not in this PR)

The base headroom on this test is at 87% on a quiet box (95s / 120s) — that's structurally fragile even with the bump. Worth a separate issue to audit which imports under createOpenClawTools are heaviest and can be lazy-loaded. Filing that as a follow-up issue, not coupled to this fix.

Provenance

Verified by a fresh subagent run (silas-pr485-fixup-v2, 15m26s, 2026-05-01 07:29 UTC) after a prior runId-thread hypothesis was disproved when an earlier subagent surfaced "the timeout happens on the FIRST test regardless of which one runs first" — pointing at module-load cost rather than feature-gated logic.

Cohort cross-eye welcome on the patch shape.